Best AI Models for the RX 9070 XT
The 16GB list on AMD’s RDNA 4 card: which backend Ollama uses on Linux and on Windows, what a second card adds, and how it compares with the RTX 5060 Ti. Footprints from Ollama’s tag sizes, speeds as bandwidth ceilings.
The best AI model for the RX 9070 XT is Qwen 3 14B — 9.3GB at Q4_K_M as Ollama lists it, ~6–7GB of the 16GB left for context, with a bandwidth ceiling of ~69 tok/s (AMD’s 640 GB/s ÷ 9.3GB — a ceiling, not a measurement). The rest of the 16GB list carries over from NVIDIA unchanged, because GGUF files are vendor-neutral: Qwen 2.5 Coder 14B (9.0GB) for code, gpt-oss:20b (14GB, MoE) as the fast assistant, Gemma 4 12B at Q8 (13GB) for quality and Devstral Small 2 24B (15GB) as the strongest coder that fits. What to know before buying is the platform split: Ollama lists the 9070 XT (gfx1201) for ROCm on Linux, but its Windows ROCm list stops at RDNA 3, so on Windows Ollama runs this card through its Vulkan backend. We do not own a 9070 XT; nothing here is a measured rate.
Models that run well on the RX 9070 XT
| Model | Size | Build | Memory used | Ceiling tok/s* |
|---|---|---|---|---|
| WINNERQwen 3 14B The 16GB winner: the whole 14B class with ~6–7GB of KV budget. The q8_0 tag is 16GB, the entire card, so the Q4 build is the one to pull. ollama pull qwen3:14b | 14B (dense) | Q4_K_M | ~9.3GB | ceiling ~69 tok/s (640 ÷ 9.3GB) |
| Qwen 2.5 Coder 14B The everyday coding model, with ~7GB of context budget for whole files: the same file and the same fit as on a GeForce card. ollama pull qwen2.5-coder:14b | 14B (dense) | Q4_K_M | ~9.0GB | ceiling ~71 tok/s (640 ÷ 9.0GB) |
| gpt-oss:20b OpenAI’s open-weight model sized for 16GB. Only ~3.6B parameters are read per token, which makes it the fast general assistant here; it fills most of the memory, so run it alone. ollama pull gpt-oss:20b | 20.9B (MoE, ~3.6B active) | MXFP4 | ~14GB | MoE — ~3.6B active per token; bandwidth is not the limit; not measured |
| Gemma 4 12B Google’s dense 12B at Q8, with image input and ~3GB spare. Drop to the 8.0GB q4_K_M build when you want the memory for context instead. ollama pull gemma4:12b-it-q8_0 | 12B (dense) | Q8_0 | ~13GB | ceiling ~49 tok/s (640 ÷ 13GB) |
| Devstral Small 2 24B The strongest agentic coder that fits, with about a gigabyte of KV room. Set num_ctx on purpose; a 32K agent loop will spill to system RAM. ollama pull devstral-small-2:24b | 24B (dense) | Q4_K_M | ~15GB | ceiling ~43 tok/s (640 ÷ 15GB) |
| Flux.1 dev FP8 (ComfyUI on ROCm) AMD’s spec page lists FP8 matrix throughput for this card (the RDNA 3 RX 7900 XTX’s page lists none), so the FP8 file is more than a memory saving here; whether your ComfyUI and PyTorch versions use that path is version-dependent, and we have not timed it. Linux uses the ROCm PyTorch wheels; on Windows, ComfyUI’s install docs point to AMD’s ROCm 10.0 PyTorch packages with a gfx1201 build. ComfyUI (ROCm PyTorch): flux1-dev-fp8.safetensors | 12B | FP8 | ~11GB (~13GB with LoRAs) | diffusion — compute-bound, not timed |
* Ceiling tok/s = the device’s vendor-listed memory bandwidth ÷ the model’s file size as Ollama lists it — the rate a dense model cannot exceed, because every generated token reads the whole weight set once. We have not measured this hardware. Real-world generation usually lands at 50–80% of the ceiling and falls as the context grows; mixture-of-experts rows read only their active parameters, so no ceiling is quoted for them, and image models are compute-bound, so no render time is.
What won't fit on the RX 9070 XT
- ✗Qwen3.6 27B (needs 17GB at Q4_K_M) — A gigabyte over before any context. The 24GB RX 7900 XTX holds it with ~7GB to spare, and so does a pair of 9070 XTs.
- ✗Qwen3-Coder 30B (needs 19GB at Q4_K_M) — The MoE coder that anchors the 24GB Radeon’s coding list. On 16GB it spills to system RAM.
- ✗Qwen 2.5 Coder 32B / Gemma 4 31B (needs 20GB each) — The dense 30B-class models are 24GB-card, or two-card, territory.
- ✗Llama 3.3 70B (needs 43GB at Q4_K_M, 26GB even at Q2_K) — Not a 16GB model at any usable quant; even two 9070 XTs (32GB) only reach it at 2-bit.
RX 9070 XT-specific notes
- →Where this page’s numbers come from: we do not own a 9070 XT. Footprints are the file sizes Ollama prints on each tag page (read 20 Nov 2026), and every tok/s is a ceiling — the “up to 640 GB/s” on AMD’s spec page divided by the gigabytes a dense model reads per token. The 64 MB Infinity Cache does not change that arithmetic for decode, because a 9GB weight set streams through it rather than living in it. We quote no measured rate; our 9070 XT guide collects third-party benchmarks with their sources.
- →Linux, stated exactly: Ollama’s GPU docs list the RX 9070 XT in their Linux ROCm table and map it to LLVM target gfx1201, with the AMD ROCm v7 driver required (AMD’s amdgpu-install utility installs it). Ollama’s install script fetches its ROCm package automatically when it sees an AMD GPU. AMD’s ROCm 7.14.0 system-requirements page lists the card (RDNA 4, gfx1201) as supported on Ubuntu 24.04.4 and 22.04.5 and RHEL 10.1 and 9.7, and it is on the compatibility matrix for ROCm 10.2.0, released 18 November 2026. No HSA_OVERRIDE_GFX_VERSION is needed; that variable is for cards that are not on the list.
- →Windows, stated exactly: Ollama’s Windows ROCm table covers RDNA 3 cards only (7900 XTX down to 7600), so the 9070 XT is not on it. Ollama’s Windows docs name “a Vulkan-capable AMD Radeon driver for Vulkan acceleration” as the other route, and Vulkan is enabled by default, so on Windows Ollama runs this card through Vulkan. The model list is the same; if you want ROCm specifically, run Linux. On either OS, `ollama ps` must read 100% GPU after a model loads.
- →Variant check: the cheaper RX 9070 (non-XT) has the same 16GB and the same “up to 640 GB/s” on AMD’s spec page, so every row and ceiling here applies to it. It gives up compute (56 vs 64 compute units, 220 vs 304 W board power in AMD’s launch table), which shows in prompt processing and image generation rather than decode. The RX 9070 GRE is a different tier: 12GB on a 192-bit bus at up to 432 GB/s, so its list is the 12GB one on our RTX 5070 page, at a lower ceiling.
- →RX 9070 XT vs RTX 5060 Ti 16GB, the cross-shop this card usually gets: identical 16GB model list, 640 vs 448 GB/s, so the 9070 XT’s ceiling is 43% higher on every dense row. The 5060 Ti answers with CUDA, 180 W instead of 304 W, and FP8/FP4 tensor cores. On 20 Nov 2026 their used prices were close (see the price note), which makes it a software decision: Ollama, llama.cpp and ComfyUI run on both, CUDA-only projects on one.
- →Two cards: Ollama’s FAQ says a model that does not fit on one GPU is spread across all of them, so a pair of 9070 XTs (32GB) holds Qwen3.6 27B at Q4 (17GB), Gemma 4 31B and the 32B coders (20GB each). It does not add decode speed for one conversation: each token still reads every weight once, part from each card in turn, so a 17GB model’s ceiling stays near 640 ÷ 17 ≈ 38 tok/s. ROCR_VISIBLE_DEVICES chooses which Radeons Ollama sees on Linux.
- →Power, from AMD’s spec page: 304 W typical board power, a 750 W minimum PSU recommendation and 2× 8-pin connectors in AMD’s specification (partner boards vary). Decode is a sustained load measured in minutes; size the supply for that, and for 608 W of board power if you add a second card.
- →Price on 20 Nov 2026: AMD’s SEP was $599 (cards on sale from 6 March 2025). videocardprices.com showed $1,034.99 at Amazon as the lowest in-stock new price, 73% over SEP, and resaleprices.com put the used average at $889 (fair asking range $863–937, 58 eBay listings). That is about what the 16GB RTX 5060 Ti listings asked the same day ($850–880) for 43% more bandwidth, and far above a used Arc A770 16GB ($349), which holds the same list at 560 GB/s with no named Ollama support.
Quick start
curl -fsSL https://ollama.com/install.sh | shrocminfo | grep gfx1201ollama run qwen3:14bollama psGo from "it runs" to actually building
All 561 chapters — running local models, RAG, agents, fine-tuning — plus the Python Lab and every course added later.
Frequently asked questions
What is the best AI model for an RX 9070 XT?
Qwen 3 14B at Q4_K_M — 9.3GB as Ollama lists it, ~6–7GB of the card left for context, and a bandwidth ceiling of ~69 tok/s (640 ÷ 9.3GB). For code, Qwen 2.5 Coder 14B (9.0GB) day to day or Devstral Small 2 24B (15GB) for agent runs; for a fast assistant, gpt-oss:20b (14GB); for quality, Gemma 4 12B at Q8 (13GB). Every figure is a tag size or a ceiling, not a rate we measured.
Does Ollama support the RX 9070 XT?
Yes: by name on Linux, through Vulkan on Windows. Ollama’s GPU documentation lists the 9070 XT in its Linux ROCm table (LLVM target gfx1201) and requires the AMD ROCm v7 driver there; AMD’s ROCm requirements list the card as supported on Ubuntu 24.04 and 22.04 and RHEL 10.1 and 9.7. Ollama’s Windows ROCm table stops at RDNA 3, so on Windows the card runs on Ollama’s Vulkan backend, which is on by default with a Vulkan-capable Radeon driver. Either way, check that `ollama ps` reads 100% GPU after loading a model.
RX 9070 XT vs RTX 5060 Ti 16GB for local AI?
Same 16GB, same model list: everything on this page runs on both. The 9070 XT has 640 GB/s to the 5060 Ti’s 448, so its decode ceiling is 43% higher on every dense model (~69 vs ~48 tok/s on Qwen 3 14B). The 5060 Ti has CUDA, draws 180 W to the 9070 XT’s 304 W, and has FP8 and FP4 tensor cores. At the near-identical used prices of 20 Nov 2026, the 9070 XT gives more bandwidth for the money if your tools are Ollama, llama.cpp and ComfyUI; if you need CUDA-only software, buy the 5060 Ti.
Is a dual RX 9070 XT setup worth it for local AI?
Only if you need the capacity. Two cards give 32GB, and Ollama spreads a model across both when it does not fit on one, so Qwen3.6 27B (17GB), Gemma 4 31B and the 32B coders (20GB each) become possible. Decode speed for a single conversation does not double: each token still reads every weight once, so a 17GB model’s ceiling stays near 640 ÷ 17 ≈ 38 tok/s. Compare the pair against one 24GB card: our RX 7900 XTX page covers the same 27–32B models at 960 GB/s on a single board.
Do I need HSA_OVERRIDE_GFX_VERSION for the RX 9070 XT?
No. That variable tells ROCm to treat an unsupported Radeon as a similar supported target. The 9070 XT is gfx1201, which is on Ollama’s Linux list and AMD’s ROCm support list, so set nothing. If Ollama does not see the card on Linux, check that the ROCm v7 driver is installed and, on some distributions, that the ollama user is in the render group; on Windows, update the Radeon driver, since Ollama uses Vulkan there.
Related guides
Other GPUs
Ready to Go Beyond Tutorials?
25 structured courses with hands-on chapters - build RAG chatbots, AI agents, and ML pipelines on your own hardware.
Was this helpful?